GNSS simulations are an important toolset supporting system engineering trade-offs and decisions. They are used to display and monitor key system performance indicators as well as study system behaviour, both through synthetic scenarios and real data replay.
Industry and ESA currently use different tools to simulate GNSS systems, however often constrained to undertake only one of two different options: i) either an accurate, but computationally intensive and rather cumbersome approach is adopted to operate simulation chains representing each part of the system separately; ii) or a much faster macro-model based tool is used, which has to be tuned and calibrated manually, however limiting its usability only to tuned scenarios.

Limitations of current macro model based tools can be overcome by training a machine-learning (ML) algorithm to predict the behaviour of the core of selected GNSS systems.
Trained on real data recorded during years of operation, the created ML tools shall be an intermediary product between the two above simulation approaches. With the user-friendly operability and low computational cost of current macro model tools, it shall aim at offering more accurate results on a wider range of different scenarios than a macro model, which otherwise would represent accurately only tuned scenarios.

Aside from absolute system performance, another important aspect in design decisions is system sensitivity to single factors, e.g. sensitivity to monitoring station positions. Advantages of using machine learning are the offered data analysis capabilities.
The main objective of the proposed activity is to assess usability of machine learning (ML) techniques in support of GNSS system simulations by modelling selected system segments using ML algorithms.
The tasks to be performed will include:

•  review of state-of-the-art in the field of Supervised Learning techniques for Multi-Target prediction, trading-off different ML techniques;
•  training of the machine learning algorithms selected during trade-off analyses on observables as input and final orbit products as output data for at least two GNSS cases (e.g. IGS, augmentation systems or others);
•  assessment of algorithm’s performance on previously selected test data sets;
•  identification of main issues in using ML to model GNSS behaviour.

The results of the activity will provide:

•  demonstration of usability of ML to model segments of a GNSS system;
•  implementation of selected ML techniques into tools to simulate GNSS system segments capabilities and behaviour.